01Strategy / The commercial lens
Where does it change your advantage?
Start with one repeated job: preparing a first draft, finding an answer in approved documents, or sorting incoming enquiries. Establish the current cost and error rate. A subscription price is only one line in the calculation; add integration, staff training, human review and correction.
Then ask the harder question: if delivery becomes faster, will you keep the benefit, pass it to the customer, or face a lower market price? That is a board discussion, not a software feature.
02Risk / The accountability lens
Who is answerable when it fails?
Section 174 of the Companies Act requires directors to exercise reasonable care, skill and diligence. AI does not replace that duty. Our practical recommendation is to record why you approved a use, what evidence you considered and who can stop it. Companies Act, s174 ↗
Ask whether the system merely drafts, recommends, or actually acts. A draft email and an agent authorised to send offers are different decisions. Give one person ownership and define when a human must approve an action.
03Technology / The control lens
What can the tool see and do?
Map the information going in and the permissions coming out. The NCSC warns that language models do not enforce a reliable boundary between instructions and data; hostile instructions can arrive inside material the system reads. Ask your supplier how access, approvals and incident recovery limit the damage. NCSC on prompt injection ↗
For personal information, check purpose, lawful basis, transparency, retention and individual rights. The ICO flags its AI guidance as under review after the Data (Use and Access) Act; its current roadmap schedules the updated automated-decision guidance for winter 2026. Check the current position before a consequential deployment. ICO AI guidance ↗ Update timetable ↗
04Workplace / The people lens
Which tasks change, and who gets support?
List the tasks before you redraw the organisation chart. Someone still needs to handle exceptions, verify results and explain decisions to customers. Ask staff what they already use and where they need training; a policy that nobody can follow will tell you very little.
The ILO’s 2025 research measures potential occupational exposure to generative AI, rather than observed redundancies. It points to job transformation as the likelier outcome. Use that distinction when planning your own pilot: test the work before assuming a headcount saving. ILO research ↗